⚡ Quick Summary
Published by the European Union Agency for Cybersecurity (ENISA), this report examines artificial intelligence as both a cybersecurity capability and a source of cybersecurity risk. It frames AI as a dual-use technology and distinguishes between “AI to protect”, including AI use cases in cybersecurity, and “AI to harm”, including malicious use and attacks against AI-based mechanisms. Its stated aim is to identify research needs for AI for cybersecurity and for securing AI under ENISA’s Cybersecurity Act mandate.
The report surveys traditional machine learning, neural networks and deep learning; links these methods to cybersecurity functions such as intrusion, malware and anomaly detection; and discusses prevention and detection use cases. It also considers security-by-design, privacy-, explainability-, robustness- and fairness-by-design, alongside threats including data poisoning, adversarial attacks, model stealing, model inversion and membership inference. Its central deliverable is five proposed research needs: AI/ML cybersecurity test beds, AI/ML penetration testing, standardised privacy and confidentiality frameworks, practitioner training using real-world scenarios, and an AI and cybersecurity threats observatory.
🧩 What’s Covered
The report proceeds from technical foundations to applications, security concerns, case studies and funding-oriented research proposals.
- Definitions and AI methods: Defines AI, AI systems, machine learning, deep learning, supervised and unsupervised ML, reinforcement learning and security-by-design. It separates traditional ML methods—decision trees, SVMs, Naive Bayes, K-means, hidden Markov models and genetic algorithms—from neural-network approaches including ANNs, CNNs, RNNs, autoencoders, Siamese neural networks and ensemble methods.
- Cybersecurity functions and data: Maps AI techniques to intrusion detection, malware detection and classification, vulnerability assessment, spam filtering, anomaly detection, phishing detection, traffic analysis, data compression and feature extraction. It lists widely used datasets including KDD Cup 99, DEFCON, CTU-13, CICIDS2017, CICAndMal2017 and IoT-23.
- AI for cybersecurity: Discusses prevention through vulnerability assessment, fuzzing, penetration testing, risk scoring and identity/access analysis. It then covers detection of spam, malware, intrusions and attacks, noting that models trained on prior incidents may fail to identify novel attack types.
- Securing AI: Sets out security-by-design practices: risk assessment and threat modelling, secure coding, secure data handling, testing and validation, and transparent and explainable design. It describes privacy-, explainability-, robustness- and fairness-by-design as AI-specific considerations.
- AI-powered attacks and defences: Describes GAN-enabled deep fakes, spear phishing, adaptive malware, training-data backdoors, software-library vulnerabilities, data poisoning, adversarial examples and reverse engineering of models. Proposed protective approaches include securing software packages and verifying training-data validity.
- Selected case studies: Examines next-generation telecommunications, IoT and IoE, cyber-physical systems and cyber biosecurity. Across these domains it stresses new attack surfaces, the need for reference data, and the challenge of adapting AI cybersecurity tools to unfamiliar contexts.
- Gaps, needs and next steps: Identifies shortcomings around data quality, bias, non-stationarity, real-world performance, latency, privacy and standardised evaluation. It translates these into five research needs and states ENISA’s planned work on priorities, a roadmap and an observatory.
💡 Why it matters?
The report helps cybersecurity and AI governance teams treat AI as both a defensive component and an attack surface. It connects familiar operational tasks—vulnerability assessment, penetration testing, anomaly detection and incident triage—to model-specific risks such as poisoned training data, adversarial inputs and model extraction. Its research-gap analysis also makes clear that accuracy alone is insufficient: reliable deployment depends on data quality, robustness, explainability, privacy, computational constraints and performance in real-world rather than simulated conditions.
For EU research and funding stakeholders, the five proposed needs provide a structured agenda spanning testing environments, privacy and confidentiality assessment, practitioner capability and threat information sharing.
❓ What’s Missing
The document is an initial assessment and explicitly says its examples of AI cybersecurity use cases are non-exhaustive. It does not provide an implementation playbook, a control-by-control audit method, or detailed technical specifications for the proposed test beds, privacy frameworks, training models or observatory. The research needs are not prioritised, and the report directs readers to a separate ENISA Research and Innovation Brief Report for the 2022 priorities. Although it discusses response and recovery as cybersecurity functions, its detailed use-case treatment concentrates on prevention and detection. The report also acknowledges that existing dataset reliability criteria focus on intrusion detection, while comparable requirements for other cybersecurity applications remain to be addressed.
👥 Best For
Cybersecurity research leads, application developers and funding bodies developing AI-related research agendas will find the research needs and proposed beneficiaries most directly relevant. It also serves security practitioners and AI system teams assessing how model choice, training data, adversarial threats, privacy and human escalation affect cybersecurity deployments, particularly in IoT and cyber-physical contexts.
📄 Source Details
Artificial Intelligence and Cybersecurity Research is an English-language ENISA Research and Innovation Brief dated June 2023. It is published by the European Union Agency for Cybersecurity (ENISA) and names Dr. Stavros NTALAMPIRAS, Dr. Gianluca MISURACA and Dr. Pierre Rossel as authors. The complete 40-page PDF gives ISBN 978-92-9204-637-8, DOI 10.2824/808362 and reference TP-07-23-202-EN-N.
